Technology

How we build.

The tools below are the ones we use on our own work today. Labels mark what is in regular use and what is working practice, so you can tell the difference.

Python and analytical workflows

In regular use

Python is the main working language for data preparation, modeling, evaluation and tooling. Analyses are written as repeatable scripts with tests rather than one-off notebooks.

Machine learning and statistical modeling

In regular use

Count and probability models, regularized regression and tree-based approaches, assessed with calibration checks and held-out evaluation. A simpler model stays in place unless a complex one demonstrably does better.

Data pipelines

In regular use

Multi-source ingestion, identity matching, validation and timestamped snapshots, stored in relational databases and object storage so a past result can be reproduced.

Cloud infrastructure and AWS

In regular use

Serverless functions, object storage, scheduled jobs and managed authentication on AWS for internal tooling. Private systems stay private, and public sites are kept separate from them.

Automation and event-driven systems

In regular use

Jobs triggered by schedules or upstream events that produce reports, grade earlier outputs and raise a flag when a dependency is stale or missing.

Data visualization

In regular use

Dashboards and reports in Python and the browser, designed to show uncertainty and provenance as clearly as the headline number.

Responsible AI integration

Working practice

AI tools assist development and analysis, but outputs are verified, claims are traced to evidence, and automated systems are built to stop rather than guess. Accountability stays with a person.

Why we test

A forecast is only as good as its track record.

Reliability diagramStated probability on the horizontal axis against how often the event actually happened on the vertical axis. A well calibrated model follows the diagonal; an overconfident one bends away from it at the extremes.STATED PROBABILITYHOW OFTEN IT HAPPENED
Interactive

Confident is not the same as correct.

Two synthetic forecasters predict the same 2,000 events. One states honest probabilities. The other sounds more certain than the evidence allows. Both look plausible until you check them against outcomes.

  • Calibrated forecaster
  • Overconfident forecaster

A note on claims

We describe what we have actually done.

Our technology experience comes from building and operating our own research systems. We do not list tools we have only read about, and we will say so when a problem sits outside what we have done before.

Have a problem worth measuring?

Tell us what you are trying to understand, automate or build. We will tell you plainly whether we are a good fit.

Start a conversation